def loadData(fname):
"""
Read the feature vector file.
Convert labels from {-1,+1} to {0,1}.
"""
labels, vects = convertSEQFileToArray(fname)
convLabels = [(y + 1) / 2 for y in labels]
return (convLabels, vects)
def train(self, trainFileName):
"""
This function implements the main training routine
for Pegasos algorithm.
"""
# Print ths stats.
print "L2 parameter =", self.l
print "Total iterations =", self.T
# Load the feature vectors.
(labels, vects) = convertSEQFileToArray(trainFileName)
# Get the (max) dimensionality d of vectors.
d = vects[0].size
self.d = d
# Get the total no. of training vectors.
n = len(vects)
print "Total no. of training instances = %d" % n
# Construct the initial weight vector.
factor = 1.0 / sqrt(d * self.l)
w = factor * ones(d)
b = 0
# Iterative procedure.
for t in xrange(1, (self.T + 1)):
startTime = time()
z = zeros(d)
bSum = 0
k = 0
for i in xrange(0, n):
x = vects[i]
y = labels[i]
score = (y * (dot(w, x) + b))
#print score
if score < 0:
# Misclassification.
z += (y * x)
bSum += y
k += 1
eta = 1.0 / (self.l * t)
whalf = ((1.0 - (eta * self.l)) * w) + ((eta / k) * z)
scale = sqrt(self.l) * sqrt(dot(whalf, whalf))
scale = min((1.0, (1.0 / scale)))
w = scale * whalf
b = b + ((eta / k) * bSum)
endTime = time()
accuracy = 100 * (1.0 - (float(k) / n))
print "Iteration: %d" % t
print "Time taken for this iteration %.3fs" % (endTime - startTime)
print "Weight norm = %f" % (sqrt(dot(w, w)))
print "Training accuracy = %f\n" % accuracy
# Set the learnt weight vector.
self.lw = w
self.bias = b
pass
def train(self, trainFileName):
"""
This function implements the main training routine
for Pegasos algorithm.
"""
# Print ths stats.
print "L2 parameter =", self.l
print "Total iterations =", self.T
# Load the feature vectors.
(labels, vects) = convertSEQFileToArray(trainFileName)
# Get the (max) dimensionality d of vectors.
d = vects[0].size
self.d = d
# Get the total no. of training vectors.
n = len(vects)
print "Total no. of training instances = %d" % n
# Construct the initial weight vector.
factor = 1.0 / sqrt(d * self.l)
w = factor * ones(d)
b = 0
# Iterative procedure.
for t in xrange(1, (self.T + 1)):
startTime = time()
z = zeros(d)
bSum = 0
k = 0
for i in xrange(0,n):
x = vects[i]
y = labels[i]
score = (y * (dot(w,x) + b))
#print score
if score < 0:
# Misclassification.
z += (y * x)
bSum += y
k += 1
eta = 1.0 / (self.l * t)
whalf = ((1.0 - (eta * self.l)) * w) + ((eta / k) * z)
scale = sqrt(self.l) * sqrt(dot(whalf,whalf))
scale = min((1.0, (1.0 / scale)))
w = scale * whalf
b = b + ((eta / k) * bSum)
endTime = time()
accuracy = 100 * (1.0 - (float(k) / n))
print "Iteration: %d" % t
print "Time taken for this iteration %.3fs" % (endTime - startTime)
print "Weight norm = %f" % (sqrt(dot(w,w)))
print "Training accuracy = %f\n" % accuracy
# Set the learnt weight vector.
self.lw = w
self.bias = b
pass
